Does Qwen3.8 27B UD-Q2_K_XL fit in the NVIDIA RTX 3090's memory?

Estimate: the UD-Q2_K_XL file is 9.8 GB and needs about 11.3 GB with an ~8K context; the NVIDIA RTX 3090 has about 23.5 GB usable, so it fits.

Qwen3.8 27B › NVIDIA RTX 3090

Qwen3.8 27B on the NVIDIA RTX 3090

Qwen3.8 27B runs at about 52.4 tokens/s on the NVIDIA RTX 3090 (measured, UD-Q2_K_XL, CUDA, single-stream llama-server decode).

Best measured: 52.4 tokens/s generation (UD-Q2_K_XL, CUDA). The source gives no prompt-processing figure.

For comparison, our size-scaling estimate for this combination is 66.7 tok/s.

Every run

Quant	Backend	Generate tok/s	Prompt tok/s	Test	Build / by / notes	Source
UD-Q2_K_XL	CUDA	52.4	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. Overclocked card.	qwen38-mtp community table (GitHub)
Q4_K_M	CUDA	31.9	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. 250 W power limit. File size 17,106,775,008 bytes as stated by the source.	qwen38-mtp community table (GitHub)
Q4_K_M	CUDA	31.9	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. 250 W power limit. File size 17,106,775,008 bytes as stated by the source.	qwen38-mtp community table (GitHub)
UD-Q4_K_XL	CUDA FA on	36.0	1,221	tg128	ee4c505 changtimwu 1x RTX 3090 on vast.ai, -fa on -ctk q8_0 -ctv q8_0.	changtimwu/vastai-exp issue #11 (GitHub)
UD-Q4_K_XL	CUDA	34.0	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. JSON/tool-call prompts.	qwen38-mtp community table (GitHub)
UD-Q4_K_XL	CUDA	34.0	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. JSON/tool-call prompts.	qwen38-mtp community table (GitHub)
UD-Q4_K_XL	CUDA	34.0	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. JSON/tool-call prompts.	qwen38-mtp community table (GitHub)
Q4_K_M		31.0	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. File size 17,106,775,008 bytes as stated by the source.	qwen38-mtp community table (GitHub)

tg128 = generating 128 tokens; pp512 = reading a 512-token prompt (llama-bench). LocalScore averages several prompt and output lengths, so its numbers run lower than tg128. "decode, llama-server" = single-stream generation speed measured against a running llama-server (no speculative decoding).

Check prices

Amazon Newegg

As an Amazon Associate I earn from qualifying purchases.

No card yet?

Try it on a rented GPU (RunPod)

Affiliate link: if you sign up, RigBench may earn a referral credit. It never affects rankings or numbers.

All results for the NVIDIA RTX 3090 · Qwen3.8 27B on other hardware

Embed this result

HTML

<a href="https://rigbench.pages.dev/model/qwen3-8-27b/on/nvidia-rtx-3090/"><img src="https://rigbench.pages.dev/badge/qwen3-8-27b--nvidia-rtx-3090.svg" alt="Qwen3.8 27B · NVIDIA RTX 3090: 52.4 tok/s (UD-Q2_K_XL, measured)"></a>

Markdown (GitHub README, forums)

[![Qwen3.8 27B · NVIDIA RTX 3090: 52.4 tok/s (UD-Q2_K_XL, measured)](https://rigbench.pages.dev/badge/qwen3-8-27b--nvidia-rtx-3090.svg)](https://rigbench.pages.dev/model/qwen3-8-27b/on/nvidia-rtx-3090/)